IR Prompt Compiler

Second half of a two-stage toolchain:

It reads a compiled IR and writes a separate file of deliverable prompts, each bound to that IR's actual figures, failure modes, and priority weights. The IR is left byte-identical.

Core Directives

1. Never Write Into the IR

The IR is the file the user uploads to a RAG system. Directives written into it become retrieval noise, not instructions.

• The IR must be byte-identical before and after this skill runs. Verify with a checksum. If it changed, you have failed.
• Why this rule exists (measured, not theoretical). An earlier version of this skill injected directives inline. The Control Plane grew to 26.6% of the document, the first 2,313 words were stage directions before a single fact, and because the manifests quoted the IR's own vocabulary, 3 of 4 chunks matching a given invariant were slide-deck choreography rather than knowledge. Chat answers degraded accordingly.
• A RAG system treats an uploaded source as evidence, not instruction. Embedded directives may sometimes steer a generated overview, since a context window has no hard privilege boundary — but they are unreliable there and actively harmful to retrieval. A prompt pasted into an instruction field is privileged; the same text inside a source is not.

2. Sidecar Emission

Write <topic>_PROMPTS.md beside the IR in the same directory. Never any other output path, and never an edit to the IR.

3. Two Forms Per Profile

Instruction fields are short; chat windows are not. Every profile emits both:

• Compact form — keep under ~500 characters. Sized for a Customize / instruction box. This is a practical ceiling, not a documented limit; when in doubt, shorter. A 2,000-word manifest cannot be pasted into these fields at all.
• Full form. For a chat turn, a system prompt, or a tool with a large instruction field. The bound manifest goes here.

4. Contextual Semantic Binding (No Generic Boilerplate)

Never write generic rules like "explain the concepts." Bind every instruction to what is actually in this file:

• Name the specific figures to focus on, and the specific node within each.
• Name the specific failure modes and metrics that deserve standalone treatment.
• Dictate narrative progression from the causal mechanisms actually present.
• Allocate by priority. Where the IR carries [Priority: X%] tags, size each chapter's manifest in proportion. Never give a 12% chapter equal airtime with a 28% one.
• Budget by Jev scores when verdicts exist. If verdicts.json sits beside the IR, size each chapter's slide/scene budget ∝ that chapter's Jev priority_score × priority_confidence, and show the arithmetic inline in the full form (Fetch 0.8×0.75 → 6 slides). Keep score arithmetic in full forms only — compacts keep one binding detail so the ~500-char ceiling holds.
• Score-vs-tag conflict: follow the Jev score and footnote the tag (6 slides per Jev priority; IR tag reads 25%). Never silently pick one.
• No verdicts file → tags-only. Fall back to the [Priority: X%] rule exactly as before; output must equal the pre-Jev behavior.
• Degrade gracefully. Not every chapter has a figure or a Core Invariant. Where one is absent, omit that field — never invent a figure or synthesize an invariant to fill a slot.

5. Inherit the IR's Uncertainty

A faithful IR marks what it does not know. The prompts must carry those marks forward, or a truthful IR still yields a confabulated deliverable.

• Never instruct the downstream AI to state an [INFERRED] claim as fact. If it earns a slide or a scene, the prompt must require it be voiced as reconstructed, not measured.
• Never instruct it to resolve an UNKNOWN_TRANSITION. That token marks a gap in the record. Surface it as an open question.
• Preserve conflict registers. Where the IR records two values for one parameter, the prompt must name the conflict rather than silently picking one.
• Carry these into the sidecar's own Fidelity Constraints section so they apply across every profile at once.
• Voice by stated band when verdicts exist. If verdicts.json sits beside the IR, each Fidelity Constraint carries its Jev stated_probability band: ≥0.9 may be voiced as measured; ≤0.1 ([INFERRED]) must be voiced as reconstructed, never measured; mid-band entries get an explicit "surface the uncertainty, do not resolve it" instruction. No verdicts file → fall back to the [INFERRED]/UNKNOWN_TRANSITION marks exactly as before.

6. Standing Q&A Rules (Per-Topic Execution Logic)

For major or complex topics, write a rule describing how to reason about that topic — trace the query through the named figure, report the node where it fails, resolve aliases before answering. These live in the sidecar's Profile C. They are never written into the IR.

• Resolve via the ontology map when verdicts exist. If verdicts.json sits beside the IR, each Profile C rule resolves terms through the verdict ontology map (alias → canonical) and the IR's frontmatter alias table. A new_term verdict becomes an "ask, don't assert" rule: instruct the downstream AI to surface the term as unresolved rather than guessing its meaning. No verdicts file → resolve against the frontmatter alias table only, exactly as before. Never invent a mapping.

7. Idempotency & Migration

• Check first. If <topic>_PROMPTS.md already exists, update it in place; do not append a second copy.
• Detect a polluted IR. If the IR already contains inline [DOWNSTREAM ...], [LOCAL DIRECTIVE ...], or <evaluation_logic> blocks, it was processed by the old inline behavior. Offer to migrate: lift those blocks into the sidecar and restore the IR to pure knowledge. Report the word count recovered. Never migrate without asking.

Output Schema (<topic>_PROMPTS.md)

# [Domain/System Name] — Deliverable Prompts
Companion to `[topic]_IR.md`. Upload the IR. Paste from this file.

## How to use
- **Audio / Video Overview** → Customize box → **B1**
- **Chat Q&A** → paste **C** at the start of a session
- **Slide deck** → paste **A2** into a chat turn or system prompt

## Fidelity Constraints (apply to every profile)
- [Claims the IR marks [INFERRED] — voice as reconstructed, never measured]
- [UNKNOWN_TRANSITIONs — surface as open questions, never resolve]
- [Conflicting values — name the conflict, do not silently pick one]

## Profile A — Slide Deck
### A1. Compact (<500 chars)
[Structure, pacing, word cap, one binding detail]
### A2. Full — bound manifest
[Per-chapter slide budget allocated by priority; named figures and focal nodes;
 named edge cases; metric placement]

## Profile B — Video / Audio Overview
### B1. Compact (<500 chars)
[Format, pacing, cold open, one binding detail]
### B2. Full — scene running order
[Fixed scene list; each bound to a named IR state change]

## Profile C — Standing Q&A Rules
[Per-topic execution logic: which figure to trace, which alias table to resolve
against, which register arbitrates contested values. Where verdicts exist,
cite the ontology map per rule (`alias → canonical` per verdicts.json);
`new_term` verdicts become ask-don't-assert rules.]

Workflow

1. Read the IR. Inventory chapters, [Priority: X%] tags, figures (type and nodes), Core Invariants, edge cases, [INFERRED] claims, UNKNOWN_TRANSITION markers, and any conflict register. Then look for verdicts.json beside the IR: if present, inventory Jev priority scores, ontology maps, and stated bands per entry. If absent, proceed tags-only — output must equal the pre-Jev behavior.
2. Checksum the IR (sha256sum). You will re-verify at the end.
3. Check for prior state. Existing sidecar → update in place. Inline directives in the IR → offer migration (Directive 7).
4. Derive Fidelity Constraints from the step-1 inventory — voiced by stated band where verdicts exist (Directive 5), by [INFERRED]/UNKNOWN_TRANSITION marks otherwise.
5. Write the sidecar to <topic>_PROMPTS.md in the IR's directory, both forms per profile.
6. Verify and report:
   • Re-checksum the IR — must be unchanged. If not, restore it.
   • Confirm each compact form is under ~500 characters.
   • Confirm no profile instructs the downstream AI to assert an [INFERRED] claim as fact.
   • Confirm every named figure and node actually exists in the IR.
   • Confirm every budget number traces to a Jev priority score (with shown arithmetic) or, absent verdicts, to its [Priority: X%] tag.
   • Confirm every alias rule traces to an ontology verdict or the frontmatter alias table — no invented mappings.
   • Present both file paths.